diff --git a/doc/manual/develop.rst b/doc/manual/develop.rst
index ad2980eae4..9b66f874a4 100644
--- a/doc/manual/develop.rst
+++ b/doc/manual/develop.rst
@@ -180,3 +180,9 @@ To debug failing Unit tests the fixtures can be loaded with:
 
 .. _CCS Contest API specification: https://ccs-specs.icpc.io/2021-11/contest_api
 .. _OpenAPI Specification ver. 3: https://swagger.io/specification/
+
+
+Additional developer documentation
+----------------------------------
+
+  - :ref:`Scoring <scoring>`
diff --git a/doc/manual/index.rst b/doc/manual/index.rst
index 1ba568179c..a70fc6f71e 100644
--- a/doc/manual/index.rst
+++ b/doc/manual/index.rst
@@ -27,3 +27,4 @@ Appendices
    shadow
    configuration-reference
    install-language
+   scoring
diff --git a/doc/manual/scoring.rst b/doc/manual/scoring.rst
new file mode 100644
index 0000000000..49337a4c94
--- /dev/null
+++ b/doc/manual/scoring.rst
@@ -0,0 +1,35 @@
+.. _scoring:
+
+Scoring and scoreboard caching implementation
+=============================================
+
+Currently, DOMjudge only supports ICPC-style 'pass-fail' problems, where a problem is either accepted or rejected.
+In this mode, we try to follow this specification, but have some additional configuration options:
+https://ccs-specs.icpc.io/draft/ccs_system_requirements#scoring
+
+Key points:
+
+- Teams are sorted by their sortorder first; this is frequently used to group teams into actual participants and other
+  groups (e.g. company teams, non-eligible teams, staff, etc.).
+- Then teams are sorted ascending by the sum of their problem points. Each correctly solved problems scores a
+  pre-defined number of points (integer, by default 1).
+- Then teams are sorted descending by either their penalty time (or runtime, can be configured at contest level). The
+  penalty time per problem is rounded down to the nearest minute by default, but can be configured to use second
+  granularity instead.
+- If two teams have the same score, they are sorted by the time of their last accepted submission.
+
+The scoreboard is updated in real-time, which can be a performance bottleneck when there are many teams and problems, so
+we employ some caching techniques to speed it up. The scoreboard cache is implemented in two tables, `scorecache` and
+`rankcache`.
+
+The ``scorecache`` table is used to store individual scoreboard cells, i.e. information about a single team's score for
+a particular problem. Whenever a submission has completed judging, the scorecache is updated for that team and problem.
+The table holds all relevant information for the team/problem combination, both for the public and restricted audience
+(i.e. jury) who has full information during the freeze.
+
+The ``scorecache`` table is then used to compute the ``rankcache`` table, which holds the aggregated information for each
+team in order to quickly compute the rank of each team. This is especially helpful when we want to know the rank of a
+subset of teams, e.g. a single team for the single scorerow on the team's overview page. For this, we do first compute
+the teams that are definitely better than that single team based on problem points and time (either penalty time or
+runtime). Then for all teams that are tied with the current team, we apply the tie breaker (time of last accepted
+submission).